AI ad changes: read every preview diff or only the risky ones?

How closely to read the preview step when an AI assistant proposes changes to a Google Ads account is something I keep coming back to. In theory you check every diff before you confirm. In practice, after the tenth preview of negative keywords that all look fine, it's easy to just hit yes.

My view is that the review should scale with the blast radius. Negatives and copy tweaks on one campaign can get a quick skim. Anything touching bids, budgets or the conversion goal deserves a real read, because those are the changes that quietly move spend for a week before anyone notices.

How do others split it? Do you review everything, or have you settled on a rule for which changes need a closer look?

Scaling review by blast radius is right, but a skim only works when the diff is small and looks the same every time. Ten negative-keyword previews in a row is exactly where fatigue sets in. I'd read in full anything that touches money or the goal (bids, budgets, conversion actions, bidding strategy), and for everything else just check the count and the one or two riskiest lines instead of every row. And if a "negatives only" batch suddenly includes a broad match on a core term, that's the one that gets a real read, whatever category it falls under.

On the batching side, the skim rule only works if a preview contains one kind of change. If an assistant bundles negatives, a bid change and a budget change into one preview, you either read the whole thing closely or skim it all, and both are wrong. Separate previews per change type cost one extra confirm, but then "negatives only" actually means something.

A count can also hide problems. "12 negatives added" doesn't tell you whether two of them are phrase match on a term that's also a top converting query. Each line should show match type and the campaign or ad group level, not just the total.

When a batch comes back mixed, do you ask the assistant to split it, or reject it and start over?

I'd ask for a split rather than rejecting the whole batch. The good lines are usually fine, and a full rejection throws them out along with the bad ones.

On match types, one more check. A phrase negative blocks any query containing those words in that order. Say a dental clinic adds "dentist" as a phrase negative to cut junk. It also blocks "emergency dentist near me," which could be the best converting query in the account. So before confirming, run each proposed negative against the converting search terms, not just the wasteful ones.

Splitting is right, and the dentist example is the one people forget. Before you confirm a negative, check it against the converting search terms too, not just the wasteful ones. A phrase negative can quietly block queries you're already paying for.

Checking negatives against converting queries is the right move, but do it in the same match-type context the negative will actually run in. A broad-match negative can look harmless in the search term report, just a few words, and still block a converting query if it shares a common word with it. So pull the full search term list for the account before you add it, not just the top converters.

I don't think that's how broad match negatives work. A broad negative doesn't block a query just because they share one word. Every word in the negative has to appear in the query, in any order, so a broad negative like "cheap dentist" only blocks queries that contain both "cheap" and "dentist."

The real danger is a single-word negative like "dentist" or "cost," because that blocks anything containing the word. So for each negative you're thinking of adding, check which converting queries contain all of its words. The full search term list is still useful, but what matters is whether the negative's words show up together in a query, not whether a query shares one word with it.